00:07:57 · Alex Immerman, George Sivulka
[00:07:57] S1: over twelve months before ChatGPT was Yeah. Like, year and a half almost. [00:08:01] S0: Yeah. Eighteen months, I guess. And I'm playing around with it, and I'm like, you know, holy shit. This is way better than anything I'd ever seen. And I remember, you know, I had one moment playing around with it where I was like, okay, it's definitely a multitask learner. Another moment. Another... And I was like, you know, I... It's not gonna be anyone working on this stuff at Stanford that is going to invent the most important technology of the next century. And I just became obsessed with understanding how to apply it and to build what I thought would be the most important product of the next century. And at the same time, you know, and and this is kinda, you know, it's like, how do you get from that to building an AI platform for knowledge workers for financial services? A lot of my peers were... If they were smart and they were lucky, they were going into financial services jobs. You know? They would go and become an investor. And I realized that the early... Like, incredibly, the smartest people in the world, incredibly smart kids, were going and doing [00:08:54] S1: the stupidest tasks. So many tedious things. Repetition. Boom. Yep.
00:08:54 · Alex Immerman, George Sivulka
[00:08:54] S1: the stupidest tasks. So many tedious things. Repetition. Boom. Yep. [00:08:59] S0: Night after night. And I saw pain. Like, I I noticed in my friends that they they literally hated their lives. And at Stanford, maybe the only other thing that I I took from, like, the relative entrepreneurship community there was build a company where there's pain. And, you know, my my 22 year old brain said, well, there's a lot of pain here, and there's the most important technology in the world that can solve it. Like, that's where I'm going to go and build a business before anyone else catches on to it. Yeah. [00:09:27] S1: You were talking about GBT three, came out in summer of twenty twenty. Fast forward a bit, ChatGBT came out in, I think, November 2022, but not specific for financial services, right? So as you were building your product, what made you realize Hebia is going to be much better for financial services use cases?
00:09:50 · George Sivulka, Alex Immerman
[00:09:50] S0: There's a piece of the ChatGPT experience, which is very magical, where, you know, it's very good at doing a single task. You can have it write a poem, or you can have it do your kid's homework. And it feels like the world's gonna end and, like, everything's fixed and we've reached nirvana AGI mode. And then you try to start to ask it to do complex tasks, things that would require multiple steps or, like, be differentiated. And, you know, the the response is always, as a large language model, I can't do that. And I think they've now, you know, changed the way it responds. That was always the response. And if you think about why it was broken, and it was good for things that were maybe, like, fun, like creative, generative, and bad for things that were requiring serious work. There's Accuracy. It's not even just accuracy, though. There's some piece of it where it's trained only on public data. And the stuff that matters for financial services, but also for law, for all of knowledge work, is offline unstructured information. And there's a piece of it where the process that is required to work is a little bit too complex. And if you look at ChatGPT and the limitations of, like, you give it to a bunch of investors and you ask it to draft an IC memo for a company like Hebia, every single VC would get the same IC memo. [00:11:06] S1: More or less. Yeah. More...
00:11:06 · Alex Immerman, George Sivulka
[00:11:06] S1: More or less. Yeah. More... [00:11:08] S0: I mean, it's not differentiated. It doesn't have custom process, custom No alpha. No alpha. And the idea was, hey, financial services is a place where the onus of putting the right information and getting the right result and having your custom process and it being transparent, I. E. Doing knowledge work, was the most important. And if this technology was fundamentally as transformative as it should be, a tool that would fix the the lives and workflows of investors would actually go to market far faster than anything that would just, you know, be a value add in in a more ephemeral or, like, single step way. [00:11:42] S1: Yeah. I think an important piece there that you mentioned is the private data. So if I'm using Hebia, I could upload all of Andreessen Horowitz's investment memos, and with that information, you could give me a very different result than if you were just searching on the web. [00:11:57] S0: That's right. Yeah. So you could put every single former IC memo in, structure that, create a library over... Javier's got, like, an infinite effective context window. And then when you look at a new opportunity, you could filter to a variety of other opportunities that were similar and actually begin to piece out how it's different, how it's similar, and all the pieces that would constitute an Andreessen Harowitz investment memo, you could
00:11:57 · George Sivulka, Alex Immerman
[00:11:57] S0: That's right. Yeah. So you could put every single former IC memo in, structure that, create a library over... Javier's got, like, an infinite effective context window. And then when you look at a new opportunity, you could filter to a variety of other opportunities that were similar and actually begin to piece out how it's different, how it's similar, and all the pieces that would constitute an Andreessen Harowitz investment memo, you could [00:12:19] S1: start to recreate. And then with Hebia, I think if you were to compare it to ChatGPT as you just did, but many of the most commonly used AI applications today, Perplexity, Claude, maybe DeepSeek recently, they're all chat interfaces. You've taken a different approach. Maybe share a little bit more about what you've done on the interface side. I think
00:12:19 · Alex Immerman, George Sivulka
[00:12:19] S1: start to recreate. And then with Hebia, I think if you were to compare it to ChatGPT as you just did, but many of the most commonly used AI applications today, Perplexity, Claude, maybe DeepSeek recently, they're all chat interfaces. You've taken a different approach. Maybe share a little bit more about what you've done on the interface side. I think [00:12:40] S0: it's the biggest change in AI interfaces since the launch of ChatGPT, where everything became a chatbot and everyone copied that. But but Hebia allows AI to do work and show its work. And to do that, we let it really put the sources first where it builds a grid out, and that grid has every single document as a row and every single column as a prompt or some sort of agentic action. And so you could imagine that as an AI is going over data or pulling numbers or pulling different quotes from documents, and you're asking it a bunch of different questions, it actually starts to build this piece of collateral that shows its work. And that collateral is effectively an agent orchestration system, an ability for humans to manage agents every single row and every cell is an agent. And it shows the complexity of what's happening with repeated tasks by an AI agent. They're very simple and elegant in human first way. And it's almost like the mind of the AI is opened up. And you could start to change [00:13:41] S1: how it's thinking.